In many patent firms, permission to use AI tools is currently restricted to the most senior individuals. Whilst almost every firm seems to have a dedicated “AI task-force” these days to trial new software, this group is usually restricted to partners or senior associates, whilst trainees and more junior associates are denied access to AI entirely. Indeed, many partners freely admit their intention for trainees or recently qualified attorneys to never be given access to AI tools for patent work, arguing that the skills of the patent profession must be forged manually through hard graft.
On the face of it, restricting the use of AI for patent work to the most experienced patent attorneys makes sense. After all, how is a trainee to learn the skills of the trade if they can simply rely on AI from day one? Furthermore, given that AI tools also need considerable input and direction from a qualified attorney in order to produce an acceptable output, giving AI to trainees is unlikely to produce satisfactory outputs. On the other hand, this approach potentially creates serious problems for the profession, including slowing (or even halting completely) AI adoption, whilst simultaneously encouraging risky AI shadow use.
The risks of shadow use
In this Kat’s view, a likely and potentially very serious consequence of demanding that the newer members of the profession refrain from all AI for professional work, is shadow use. Blocking access to every available model is probably a technical impossibility for IT departments. After all, if access to the major platforms is restricted, individuals may simply pivot to using one of the many alternative Large Language Models (LLMs), such as Mistral, DeepSeek and Grok, to name but a few.
Any shadow AI use by attorneys in a firm is, of course,
highly problematic. First, such use is generally going to be via personal
accounts rather than secure enterprise platforms. Outside of enterprise
commercial versions, the foundational LLM providers by default retain the data
inputted into these public accounts to train and update their models. The
inputs are therefore not considered confidential (IPKat). This means that any sensitive
information fed into the prompt could equate to public disclosure. Feeding
client confidential data into these systems is likely to constitute a
severe breach of confidentiality. Even if the shadow user attempts to avoid
inputting direct client confidential data, the security risks are
huge. The problem simply cannot be ignored.
 |
| Shadow use |
Added to the risk of shadow use is the problem that restricting access to AI systems is likely to hinder AI adoption and development within a firm. This Kat suspects that restricting AI usage to only the partners within a firm, in many cases, may mask an underlying reluctance for AI adoption. Senior members of a firm may have little incentive to embrace new technology that might threaten the traditional revenue stream of the firm, based as it is on the billable hour (IPKat). After all, many of the "proud dinosaurs" freely admit they are simply hoping to retire before they are forced to adapt. These attorneys are the ones who will regale you with a humorous anecdote about a severe hallucination they experienced a couple of years ago. After trying an AI-wrapper for patent drafting, they insist the technology is simply not ready.
By contrast, the younger members of the profession are uniquely positioned to assess and improve AI systems. Scientists entering the field from university or industry will now probably be native users of the technology who use it all the time for many tasks. This Kat can quite imagine their consternation when they enter the patent profession and are told that they must abandon all use of AI.
Limiting AI use to only the senior attorneys (many of whom will have very clear ideas of how things should be done, and why the old ways are best), is unlikely to be the best way of assessing all of the potential value of the software. Meanwhile, the individuals probably most adept at prompting and workflow optimisation are left unutilised and unable to share their knowledge (IPKat).
In this context, the greatest barrier to AI adoption within the patent industry, as this Kat sees it, is the general lack of appreciation in the profession of what these tools are capable of, when the effort is put in to use them properly (IPKat). Just as new trainees add value with their recent direct experience of the science, they can also teach the more mature members of the profession how AI can be used.
Corporate vulnerabilities
Of course, companies and clients face an identical AI crisis to patent firms. Worse still, corporate employees are often far less aware of the risks associated with inputting confidential information into personal AI accounts lacking enterprise-grade security guardrails. If a researcher inputs a new antibody sequence or chemical structure into a public chat interface, they risk this information being considered a public disclosure and potentially invalidating a future patent. In-house IP departments therefore have an important role in educating company employees about the risks of AI use outside of the confidentiality guardrails. Using these systems is just too easy.
Of course, the most
effective solution to this problem would be to provide enterprise-grade access across the
organisation. However, this can be prohibitively expensive. Companies must weigh up the financial costs of enterprise-grade AI
for everyone, versus the risks of inevitable shadow use in its absence.
Whatever decision is made, a comprehensive AI policy and practical AI training are essential. Patent attorneys should now be equipped with the knowledge to advise their clients on this issue. This means understanding the difference in confidentiality provisions between personal accounts and those with enterprise-grade security, the different levels of confidentiality available between different enterprise accounts, the confidentiality and security risks of AI-wrapper software and how to assess this, the data storage provisions of the AI tools you and your clients are using and how this may affect legal privilege and discovery during litigation, and what workflows are in place to validate AI outputs.
Final thoughts
The profession cannot afford to bury its head in the sand
when it comes to AI. In this Kat’s view, banning the technology outright will
only push its use into the shadows where the greatest dangers lie. Even
ignoring these risks, not allowing the new entrants to the profession to use
and help develop the AI tools employed within a firm will severely hamper AI
adoption.
The problem faced by patent firms is similar to that of
universities. Some universities have an outright ban on the use of AI by
students, whilst others require its use. As readers might suspect, this Kat is
all in favour of giving everyone access to AI, and sees no reason why this
should be an impediment to effective learning or training. Every work-product
will still need review and sign-off by a qualified attorney. Given how much we
still need the human-in-the-loop when it comes to the use of AI in patent work, trainees using AI who do not know the IP law will not produce something that will
pass muster. Using AI will simply speed up the process of generating their
first draft, and may even be an effective education tool. As with all AI usage
in the patent industry, the key is knowing how to use these tools
effectively.
Further reading
- The spectre of hallucination
- Solving the confidentiality problem
- Are you behind the wheel or waiting for the bus?
- Are AI-generated materials legally privileged? United States v. Heppner
- The future of the patent profession: Are we looking into an AI abyss?
- Is AI software for IP just expensive wrapping paper?
- AI in the patent industry: Don't believe the hype. Believe the data.
Reviewed by Dr Rose Hughes
on
Monday, May 18, 2026
Rating:
Follow the IPKat on LinkedIn 
I am largely repeating what I said in my comment to the last post on this topic (April 15, 2026), but there is a topic which I think isn't currently being discussed enough: economics.
ReplyDeleteMore specifically, the economics of the AI tools themselves, i.e. what they cost to the seller (here, whatever AI service provider a patent law firm is buying from), and ultimately what cost is passed over to the buyer (here, a patent law firm).
As things stand, AI is nowhere near profitable for any seller. OpenAI and Anthropic lose incredible amounts of money, to the tune of billions of US dollars per quarter, if not tens of billions. "Hyperscalers" (or whatever you want to call them) also lose money. "AI labs" (or whatever you want to call them) also lose money. So far, venture capitalists, "private credit", and the largest tech companies (Microsoft, Google, Amazon, Oracle) have eaten the losses. But they can't afford to do so forever. At some point, the investment has to be recouped, or the whole endeavor has to stop. People like Ed Zitron are documenting this extensively.
Even more concerning, some AI providers (like Microsoft for GitHub Copilot) have already switched from subscription billing (you pay X US dollars per month) to token-based billing (you pay tokens as you use them). And the latter is 1/ vastly more expensive, to the point that some companies are on track to spend 5% or even 10% of their wage bills just on tokens, and 2/ unpredictable, to the point that some companies report having spent their entire 2026 token budget by May. And we are not talking small numbers here; we are talking millions of US dollars per month for a large tech company.
I do not have a crystal ball, but to repeat myself, I think no one should assume that all currently available AI tools will still be on the market in a few years, and at the same price plus a little inflation adjustment.
Dear Extraneous Attorney – I usually try to avoid responding to IPKat comments, as this way madness lies! However, as you have taken the trouble to post your view twice, I thought I would give you my thoughts on this.
ReplyDeleteI believe that your assumption that AI providers are losing money on each prompt is wrong. The economics of foundational models are actually best understood as an R&D model. If you look at it in isolation, inference (the actual processing of the prompts from users) is thought to be already profitable. For the foundational labs, the billions of dollars currently being burned are instead massive capital expenditures to build compute clusters, buying hundreds of thousands of GPUs, to train the next generation of models. The LLM labs therefore aren't subsidising today's usage at a loss, but are instead funding future capabilities. They are all in a highly competitive race to be the best. If they all stopped investing in this “R&D” today, their infrastructure expenses would vanish, and they could all make money by selling access to their existing models as SaaS businesses.
Additionally, the unit cost to deliver a given level of intelligence is actually falling, not rising, approx. 3x-5x annual cost reduction, even before factoring in hardware improvements. Also, if we just look at the market, over the last couple of years, foundational labs have repeatedly slashed API prices for older models while maintaining previous performance levels. As hardware and algorithmic efficiency get better, the unit cost of providing AI services is thus going down, not up. There is a question of how much better we need the models to be. As the models get better and better, the decision of which model you need to pay for, for a specific task, will become more and more important from a business perspective. The switch to token based billing usually reflects this switch to allowing users to choice a better and more expensive model.
However, I entirely agree with you that many of the VC-funded start-ups based on AI wrapper software for IP are probably doomed. We have already seen this with the AI wrapper coding companies. No one uses these anymore when you can just use Claude Code directly in your local environment instead. In IP, the market is also far smaller, and already super crowded, with little to distinguish between the different players. So basing your firm's entire AI use on one of these wrapper companies is highly risky. But I do think OpenAI and Google will still be with us in a few years.
Dear Dr. Hughes - thank you very much for taking the time to responding to my comments.
DeleteI fully agree that would be highly risky to base one's firm entire AI use on one wrapper company. I also fully agree that Google will still be with us in a few years: it makes so much money on other product lines that it can afford to burn huge sums on AI (and in fact, the same is true of Meta, apparently on an even larger scale).
As to the rest, I am very grateful that you provided me with an opposite viewpoint to consider. But I believe it's still an open question, whether the economic gains will remain for long (if they ever materialize):
If an IP firm buys an AI service, will it really matter that older models are less expensive, if it needs the newest models (which apparently burn more tokens, cost more to train, and so on) to keep up with the competition?
If an AI service needs to switch to token-based billing and increase token prices to avoid bankruptcy, will it still be worthwhile for an IP firm to keep buying? Let's say that the price switches from a 300 USD/month/user subscription, to above 1000 USD/month/user based on token usage, with possibly much higher spikes: won't the firm's CFO start raising concerns?
To be clear, I don't claim to be an expert on the topic, I don't even have the answers to the questions I just asked, and I'm not even personally invested in the success or failure of anything AI, except of course that I would very much like to continue working as a patent attorney. I am just surprised that so few people in IP seem to be looking at the long-term economics of the whole thing. (Perhaps I am wrong, in which case I would welcome being corrected!) The long-term economics matter though, not only because most actors are in there to make money, but also because the investments are so large, both in absolute terms and compared to the revenue earned so far.
Access to affordable advanced AI tools is subject to a host of uncertainties (links of big tech to the US administration, digital sovereignty issues, political blowback, energy and water supply of data centers, etc) and should not be considered a given. Firms should be wary of such risks and be cautious before making deployment decisions having irreversible implications.
ReplyDelete